Introduction

Histologic grade is an important prognostic biomarker, which is widely used to guide breast cancer treatment [1,2,3]. In the Netherlands, grade currently indicates the need for adjuvant chemotherapy in nearly a third of breast cancer patients [1, 4]. For ductal carcinoma in situ of the breast (DCIS), grade influences radiotherapy decisions [1, 5] and indicates the need for a sentinel lymph node procedure [1]. Moreover, histologic grade may become the single biomarker that decides whether DCIS patients should or should not be treated, as this is now being investigated by multiple clinical trials [6,7,8,9].

Subsequently, accurate, and reproducible grading is of major clinical importance. However, significant inter- and intra-laboratory variation in histologic grading of invasive breast cancer (IBC) and DCIS exists in daily pathology practice [4, 10]. Providing laboratories with feedback reports, in which laboratory-specific case-mix adjusted proportions per grade were benchmarked against other laboratories, resulted in a promising, yet small, decrease in grading variation [11]. As substantial differences in grade-specific proportions were observed between pathologists within individual laboratories [4, 10], we hypothesized that variation in grading may be primarily explained by differences in grading practices of individual pathologists.

Therefore, we believe that training of pathologists in the assessment of histologic grade could attribute to better synchronization, and thereby a decrease in grading variation of both IBC and DCIS. This is further supported by Elston and Ellis, who emphasize that grading should only be performed by trained pathologists [12]. Yet, in the Netherlands, pathologists or pathology residents are currently not specifically trained in histologic grading of (pre)malignant breast lesions.

Training of pathologists by e-learning seems feasible, as e-learnings have been shown to decrease grading variation of dysplasia in colorectal adenomas [13] and to improve consistency in the histopathological diagnosis of sessile serrated colorectal lesions [14]. In addition, an e-learning module is easily accessible online for medical professionals throughout the country, or even worldwide, without the need for a live tutor or planned course days [15]. Thus, it enables pathologists to train when and where they want.

The aim of this study was to evaluate whether an e-learning module may be an effective tool to decrease grading variation of (pre)malignant breast lesions (i.e., DCIS and IBC) by studying variation in grading of a digital set of lesions by pathologists before and after the e-learning.

Materials and methods

E-learning design

The e-learning module was designed by the co-authors and agreement on its content was reached by three expert breast pathologists (PvD, CV, RG) and one expert breast pathologist’ assistant (NtH). The e-learning presents the key-concepts of histologic grading of (pre)malignant breast lesions, including background information, discussion of the specific grading classifications for DCIS and IBC, and an extensive review of these criteria with example images.

The e-learning module consists of two separate modules, one on grading of IBC and one on grading of DCIS. Pathologists and residents could participate in either one or both modules. Separate analyses were conducted for the DCIS- and IBC-module.

All images in the e-learning module were derived from breast cancer cases from daily pathology practice in our institute based on consensus. All patient-related information was removed from all images in the e-learning module to comply with the General Data Protection Regulation.

Classifications of histologic grading

For grading of DCIS, we used the classification of Holland [16] in our e-learning (Supplementary 1), as this is the guideline recommended by the Dutch breast cancer guideline, and, as we know from our previous survey among pathologists [10], this seems to be the most widely used classification in daily pathology practice.

The modified Bloom and Richardson guideline (Elston–Ellis modification of Scarff–Bloom–Richardson grading system, also known as the Nottingham grading system) [17, 18] (Supplementary 2) was used in this e-learning module for IBC, as it is the most widely used grading system for IBC and it is globally incorporated in breast cancer guidelines [1, 3, 19, 20]. This classification combines the assessment of cell morphology (nuclear pleomorphism), the proliferation (mitotic count), and differentiation (tubule formation), resulting in a total score and subsequent grade [18].

Recruitment of participants

Pathologists and pathology residents were invited to participate in this e-learning through the news-bulletin of the Dutch Society for Pathology (NVvP).

Study design

Baseline grading variation

Before they could start the e-learning module, participants were obliged to assign a grade to 52 snapshots of DCIS lesions, as they would in daily practice, to determine the baseline variation in grading between participants for DCIS. As agreed upon by consensus of the expert-panel (PvD, CV, RG, NtH) 16, 19, and 17 of these snapshots were DCIS grades I, II, and III, respectively. Several examples of these snapshots can be found in Supplementary 3. All architectural patterns (solid, cribriform, papillary, and micropapillary) were present in these tests, and some cases included comedonecrosis.

For IBC, participants were obliged to score snapshots per subcategory of the modified Bloom and Richardson guideline [17, 18], i.e. tubular differentiation (19 snapshots; 7 category 1, 5 category 2, 7 category 3), nuclear pleomorphism (19 snapshots; 6 category 1, 6 category 2, 7 category 3), and mitosis (30 mitoses, 20 non-mitoses). With regard to mitoses, participants were asked to state whether they would count 50 figures as a mitosis or not. This was done because we hypothesized that recognizing mitosis is the bottleneck, rather than counting itself. Furthermore, this way, snapshots were sufficient, while whole slides are used for this in daily clinical practice (and pathologists select the field where they count). Several examples of these snapshots can be found in Supplementary 4. These snapshots were primarily derived from invasive ductal carcinoma cases and tubular, cribriform carcinomas for tubule formation, and some invasive lobular carcinoma cases. Since snapshots for tubule formation, nuclear pleomorphism, and mitoses were taken from different cases, only improvement for individual subcategories could be assessed.

After scoring the pre e-learning set of DCIS lesions and/or all images of the IBC subcategories, participants would start the e-learning itself.

Post e-learning grading variation

After completion of the e-learning, participants were asked to re-grade the same DCIS lesions, and/or IBC lesions to determine the post e-learning variation in grading between participants. All lesions were presented in a different order than prior to the e-learning. Only after this second round of grading, participants received feedback on their answers.

Outcome measures and statistical analyses

Variation in grading was measured according to several outcome measures both pre- and post-e-learning. We included scores of participants who completed grading of both the pre- and post-e-learning set of lesions, for DCIS and/or IBC. Hence, a fully-crossed design was used [21].

As primary outcome measure, the inter-rater reliability (IRR) was calculated by Light’s kappa [22], which is a type of kappa-statistic suitable for a fully-crossed study-design [21]. This overall kappa represents the arithmetic mean of kappa scores for all coder pairs [21, 22]. The 95% confidence interval for kappa was calculated using bootstrapping from 1000 replications. Interpretation of the kappa-statistics was performed according to the proposed interpretation of Cicchetti [23] (Ƙ < 0.40: poor, Ƙ 0.40–0.59: fair, Ƙ 0.60–0.74: good, Ƙ 0.75–1.00: excellent). Secondary outcome measures were the number of images scored unanimously (i.e. agreement by 90 and 100% of scoring participants) and the number of single lesions scored as both grades I and III (for DCIS) or both category one and three (for IBC: tubular differentiation, nuclear pleomorphism). Lastly, the average number of discrepancies with grade (DCIS) or sub-score (IBC) from the reference score by the expert panel (PvD, CV, RG, NtH) was determined, overall, and for subgroups of participants (expert breast pathologists, general pathologists, and residents).

All statistical analyses were performed using IBM SPSS Statistics version 25 and R [24].

Results

Participants

Thirty-six (DCIS) participants were included in the data-analyses for DCIS, while 21 participants were included in the data-analyses for IBC, as they completed both the pre- and post-e-learning grading sets for the separate modules. The majority of participants identified themselves as general pathologist, followed by expert breast pathologists, and residents (Table 1). Expert breast pathologists executed their function for the largest numbers of years, with a mean of >13 years (Table 1).

Table 1 Characteristics of e-learning participants.
Full size table

DCIS

All outcome measures for DCIS improved after the e-learning (Table 2). Overall agreement (IRR) improved from fair (kappa: 0.532) to good (kappa: 0.657) [23], and the amount of lesions graded unanimously by >90% (33/36) and 100% (36/36) of participants increased after e-learning, with best observed agreement for DCIS grade III. Interestingly, before the e-learning, almost 30% of DCIS images (n = 15) were graded as both grades I and III by the different participants, which decreased to <10% (n = 5) after the e-learning (Table 2). Lastly, the average number of discrepancies with the reference score by the expert panel decreased for all subgroups of participants.

Table 2 Variation in grading of ductal carcinoma in situ of the breast before and after e-learning (36 participants).
Full size table

Invasive breast cancer

All outcome measures that contribute to grading, except for two outcome measures for nuclear pleomorphism, improved after e-learning (Table 3).

Table 3 Variation in scoring of the 3 constituents of invasive breast cancer grading before and after e-learning (21 participants).
Full size table

For tubular differentiation, overall agreement (IRR) improved from good (kappa: 0.653) to excellent (kappa: 0.846) [23], while the number of images graded unanimously by >90% (19/21) and 100% (21/21) participants increased as well (from 63.2% to 89.5%, and from 0.0% to 47.4%, respectively). The number of images scored both as category 1 and 3 was notably high before e-learning (78.9%), yet a substantial decrease was observed after e-learning (10.5%). Finally, the average number of discrepancies decreased for all subgroups of participants.

For nuclear pleomorphism, overall agreement (IRR) remained only fair (kappa: 0.523 vs. kappa: 0.571) after the e-learning module. The number of images graded 100% unanimous slightly increased (from 15.8 to 26.3%), while the number of images graded unanimously by >90% of participants remained stable (31.6% pre- and post-e-learning). Overall, for both outcome measures, best agreement was observed for nuclear pleomorphism category 3. After e-learning, an increase was observed for the number of images scored as both category 1 and 3 (0.0–15.8%), all of which were deemed category 2 by consensus of the expert-panel. Hence, for these lesions, the same lesion may be treated differently in daily clinical practice, based upon which pathologist graded it, as this is a difference of two points on the overall score, and most probably would lead to a different overall grade (3–5 = grade I, 6–7 = grade II, 8–9 = grade III). Lastly, the average number of discrepancies did show improvement after e-learning (27.1 to 18.8%).

For mitoses, overall agreement (IRR) showed improvement from good (kappa: 0.699) to excellent (kappa: 0.826). After the e-learning, an increase was observed for the number of lesions scored unanimously by >90% and 100% of participants (from 36.0% to 56.0%, and from 74.0% to 86.0%, respectively). The overall number of discrepancies decreased from 10.5% before e-learning to 6.6% after e-learning.

A few example images with low- and high- concordance after e-learning for both DCIS and IBC can be found in Supplementary 5.

Discussion

This study shows that an e-learning, in which pathologists and residents are trained in histologic grading of DCIS and IBC, is a feasible and promising tool to decrease grading variation, as 18/20 outcome measures improved after e-learning. This is highly relevant considering the substantial grading variation of (pre)malignant breast lesions in current daily clinical practice [4, 10] bearing in mind that grading plays a decisive role in clinical decision making such as the indication for chemotherapy in IBC, and for a sentinel lymph node procedure, and partial breast irradiation in DCIS. Increased consensus with regard to grading will diminish variation in treatment and thereby most likely also patient outcome [4].

To evaluate the effect of our e-learning, we used multiple outcome measures (Tables 2 and 3). The IRR was chosen as outcome measure as this provides a way of quantifying the degree of agreement [21] between the e-learning participants on histologic grading. It is important to acknowledge that kappa-values may be influenced by the choice of the specific kappa statistic, and there is no clear guideline which specific kappa statistic to use for a fully-crossed design with multiple (≥2) coders and an ordinal outcome measure (grades I–III, category 1–3) [21, 25,26,27]. Here, we chose Light’s kappa [21], which uses the arithmetic mean of the kappa’s of all possible coder pairs.

To provide further insights, we assessed the number of lesions scored unanimously (both by all, and by >90% of participants), and the number of lesions which were scored both grades I and III or category 1 and 3, as this outlines where difficulties may and may not lay (specific grades/subcategories).

For DCIS, almost 30% of the pre-e-learning images were scored as both grades I and III (of which the majority was deemed grade II by our expert panel). Although most discrepancies were only one grade apart, this may point out a challenge for current clinical trials, who focus on the safety of active surveillance in grades I and II DCIS patients [6,7,8,9], especially since central pathology review is not carried out in all trials [28]. This means that, in current daily pathology practice, grade III DCIS lesions may erroneously be deemed grade II and randomized to active surveillance. This underlines that central pathology review is essential for these trials, and raises the question whether histologic grade, in its current state, should be the single identifying biomarker for low-risk DCIS, as is also pointed out by Cserni et al. [29].

For IBC, most difficulties were observed for scoring nuclear pleomorphism, which is in line with previous studies [4, 30,31,32,33], and may be due to the fact that scoring of this subcategory is least quantitative, leaving most room for variation in interpretation. It should also be mentioned that we used snapshots of lesions (of very similar magnification), which makes interpreting nuclear pleomorphism somewhat more difficult, especially since comparison to other cells (for example epithelium) by zooming out was not possible. In contrast, for tubular differentiation and mitoses, >90% of participants agreed on almost 90% of the questions after the e-learning, indicating that for these two constituents of grade the learning effect is much greater.

A limitation of this study may be that participants graded the same digital set of lesions after finishing the e-learning. Participants could go through the e-learning (and both tests) at their own pace, even in multiple sessions, so it is unknown what the “wash-out time” was between the pre- and post-test. Therefore, it cannot be ruled out that, although the lesions were presented in a different order, participants “recognized” some pictures. However, we would like to emphasize that participants only received extensive feedback after the final test. In addition, grading the same set of lesions, was deemed the best way to observe the effect of the e-learning on grading variation as this enables comparing “baseline” variation before e-learning with grading variation after e-learning. Grading a different set of lesions after the e-learning would not enable us to distinguish between an effect of the e-learning or simply because grading of the post-e-learning set may have be “easier” of “more difficult”.

Furthermore, as complete consensus was not always obtained after e-learning, the images chosen based on “true grade/category” consensus of our expert panel may still not be the perfect examples. Nonetheless, we would like to emphasize that the average number of discrepancies from this “true grade/category” decreased after e-learning for both DCIS and all subcategories of IBC. Therefore, if this type of training by e-learning is to be enrolled on a larger scale, one could think of a larger consensus panel of expert breast pathologists, especially since, for tubular differentiation and nuclear pleomorphism, most discrepancies from our expert panel were observed among this subgroup of participants.

A final limitation may be our decision to use mitosis identification in snapshots, rather than the selection of high power fields and actual mitosis counting. This was done because of practical reasons, however, we would like to emphasize that this indeed leaves out selection of high power fields, which may be a challenge, especially in a heterogeneous tumor, leaving room for grading variation within this subcategory.

Supported by the results of this study, we believe that training of pathologists in the assessment of histologic grade could further attribute to better synchronization and thereby a decrease in grading variation of both IBC and DCIS. An e-learning module as described in this study, in combination with monitoring variation in histologic grading of IBC and DCIS in daily clinical practice by laboratory- and pathologist specific feedback, may be considered by the Dutch Society of Pathology as one of the pathways for those pathologists that want to qualify as breast pathologist. Likewise, pathologists who participate in the Dutch national colorectal cancer screening program are already obliged to participate in an e-learning. In addition, the use of peer-learning programs for training and quality was identified as one of the strategies to improve patient access to high-quality oncologic pathology by Nass et al. [34]. Lastly, the e-learning may well be implemented in training programs of pathology residents.

While both laboratory-specific feedback and training of pathologists by e-learning may be promising tools to decrease grading variation, we believe that it is also important to emphasize that histologic grade is not a fact, nor the truth, but merely a model or tool that consists of a statistically computed set of cut-off values in a spectrum of histopathologic features to classify expected tumor behavior. Yet, these cut-offs are often used as hard criterion by clinicians in clinical decision making. For example, according to the Dutch breast cancer guideline, for a large group of patients (≥35 years, N0-, HER2-, 1.1–2 cm) [1], ≥grade II indicates that they are eligible for chemotherapy, whereas grade I does not. Yet, the majority of patients (>75%) has a score on the switch point of grades, that is scores 5 or 6, and scores 7 or 8 [4]. Thus, for these patients, the difference of only one point on any of the subcategories of the Bloom and Richardson classification may already alter their overall histologic grade, and thus may have different therapeutic implications. Therefore, we believe that awareness, and understanding of the difficulties of histologic grading, among clinicians is crucial as well to improve clinical decision making for patients. Furthermore, in the current era of shared-decision making, these difficulties should also be discussed with patients, where relevant.

In conclusion, this study shows that an e-learning module, in which pathologists and residents are trained in histologic grading of DCIS and IBC, is a feasible and promising tool to decrease grading variation of (pre)malignant breast lesions. This is highly relevant given the important role of histologic grading in clinical decision making of DCIS and IBC whereby an influence on outcome cannot be ruled out.